You are on page 1of 8

An Empirical Procedure for

Fatigue Damage Estimation in


Instrumented Risers
In order to assess the effects of vortex-induced vibration (VIV) and to ensure riser integ-
C. Shi1 rity, field monitoring campaigns are often conducted wherein the riser response is
School of Petroleum Engineering, recorded by a few data sensors distributed along the length of the riser. In this study, two
China University of Petroleum (East China), empirical techniquesproper orthogonal decomposition (POD) and weighted waveform
Qingdao, Shangdong 266580, China analysis (WWA)are sequentially applied to the data; together, they offer a novel empiri-
cal procedure for fatigue damage estimation in an instrumented riser. The procedures
are briefly described as follows: first, POD is used to extract the most energetic spatial
L. Manuel modes of the riser response from the measurements, which are defined only at the avail-
Department of Civil, Architectural, and
able sensor locations. Accordingly, a second step uses WWA to express each dominant
Environment Engineering,
POD mode as a series of riser natural modes that are continuous spatial functions
University of Texas,
defined over the entire riser length. Based on the above empirically identified modal
Austin, TX 78712
information, the riser response over the entire length is reconstructed in reversei.e.,
compose identified natural modes into the POD modes and, then, assemble all these dom-
inant POD modal response components into the derived riser response. The POD proce-
dure empirically extracts the energetic dynamic response characteristics without any
assumptions and effectively cleans the data of noisy or less important features; this fun-
damental application of WWA is used to identify dominant riser natural modesall this is
possible using the limited number of available measurements from sensor locations.
Application of the procedure is demonstrated using experimental data from the Norwe-
gian Deepwater Programme (NDP) model riser. [DOI: 10.1115/1.4035303]

Introduction challenge to estimate fatigue damage using the real-time measure-


ments. Compared with analytical approaches, fatigue damage esti-
With exploration of hydrocarbon resources moving into deeper
mated empirically using response measurements has some
waters, offshore structures are encountering harsher environmen-
advantages: (a) the inputs (measurements) are explicitly taken
tal conditions. Flow-induced vibrationsincluding VIV as well
into account of any complex characteristics of the risers dynamic
as WIVare often observed in deepwater risers due to their
response and (b) fewer or no assumptions and simplifications are
slenderness and light damping. VIV is a complex fluid-structure
needed on the riser model and on the current velocity profiles.
coupling phenomenon; its mechanism is not yet fully understood.
Thus, the development of empirical fatigue damage estimate
Sustained large-amplitude VIV is a major contributor to accumu-
methods for risers has become an important area for research in
lated fatigue damage in risers and, thus, this problem has attracted
the offshore oil and gas industry.
much practical interest in the offshore industry. In the design
Several modal identification techniques, such as POD [2],
phase for a riser, VIV and associated fatigue damage are often
weighted waveform analysis (WWA) [3,4], modal phase recon-
analyzed using finite element analysis (FEA) software, in which a
struction (MPR) [5], and efficient modal decomposition and
discretized finite element model of the riser is constructed based
reconstruction (EMDR) [6], have been employed by researchers
on various simplifications/assumptions on the risers physical and
to study a risers simulated or measured VIV response. In a previ-
hydrodynamic properties. Environmental conditions, such as inci-
ous study by Shi et al. [7], empirical procedures, each employing
dent current velocity profiles, are also simplified. Fatigue damage
a different modal analysis technique, were developed for the pur-
is then predicted based on the risers dynamic response simulated
pose of riser VIV-related fatigue damage estimation. The study
using such a simplified model.
demonstrated that the empirical procedures developed, though
In order to investigate VIV as well as to ensure the integrity of
they possessed distinct intrinsic relative advantages and disadvan-
the riser system, monitoring campaigns are often undertaken [1]
tages, could all be used to estimate fatigue damage over an entire
wherein the riser response (including bending strains and/or accel-
riser span by using measurements at a limited number of loca-
erations) is recorded by a number of sensors spatially distributed
tions. Among the empirical procedures studied, POD has special
along the riser axis. Generally, current velocity profiles are also
advantages, because (a) POD directly extracts energetic spatial
often recorded at a nearby location. Riser monitoring campaigns
modes from second-order statistics of the measurements and the
can thus generate a large number of data sets. Each data set con-
procedure is straightforward and the computation can be very fast
sists of a series of riser response time series measured at several
and (b) a small number of POD modes (termed dominant
discrete locations along with an associated current speed profile.
modes) usually account for a large portion of the total energy, and
Riser monitoring campaigns provide an opportunity as well as a
thus, discarding the remaining trivial modes (modes that account
for a relatively small portion of the total energy may be thought to
1
Corresponding author. be associated with noise) has a practical benefit, which is to
Contributed by the Ocean, Offshore, and Arctic Engineering Division of ASME
for publication in the JOURNAL OF OFFSHORE MECHANICS AND ARCTIC ENGINEERING.
clean the data. However, one of the disadvantages of the POD
Manuscript received July 1, 2016; final manuscript received November 10, 2016; procedure is that the empirically extracted modes are only defined
published online February 17, 2017. Assoc. Editor: Marcelo R. Martins. at the locations of the input sensors. In an earlier study [7], it was

Journal of Offshore Mechanics and Arctic Engineering JUNE 2017, Vol. 139 / 031701-1
C 2017 by ASME
Copyright V

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or


shown that if a risers response is desired at a location where no Table 2 The four bare riser NDP data sets
sensors are available, interpolation or extrapolation is needed. The
accuracy of such response estimation decreases with the distance Event Current Max. current Largest RMS
of the desired location from nearby sensors, which limits the no. profile speed (m/s) CF-disp/D
applicability of this POD procedure.
2120 Uniform 1.4 0.44
In the present study, a new approach aimed at improving the 2150 Uniform 1.7 0.40
empirical POD procedure is developed. The procedure is briefly 2350 Linearly sheared 0.7 0.42
described as follows. First, POD is used to decompose the original 2420 Linearly sheared 1.4 0.37
data (measured time series of the riser response) into a series of
energetic vibration patterns, i.e., the POD modes. Second, by
application of the WWA technique, each dominant POD mode
shape is further decomposed into several riser natural mode four tests are well suited for this study. Note that, in Table 2, the
shapes, which are spatially continuous over the riser span. Third, RMS CF displacement reported is based on the entire length of
the identified natural mode shape coordinates are used to approxi- record and the largest value from all the eight accelerometer log-
mate the dominant POD mode shape. Finally, the risers response gers was considered. Additional characteristics of the riser
over its entire span is obtained by summing the effects of all the response as observed in the four data sets may be found in the
dominant POD modes. The rationale for combining POD and work of Shi et al. [4].
WWA is that: (a) spatially continuous functions are desirable For a straight rier, such as a top-tensioned riser, VIV
since the empirical POD mode shapes are only defined at sparsely response in the CF direction is usually larger than that in the
distributed discrete sensor locations and (b) the dominant POD IL direction; hence, the empirical method presented is
modes are often narrow-banded, which makes it possible to apply employed to reconstruct the CF response. The accuracy of the
WWA even when only a few sensors are available. Note that the response reconstruction is examined using a leave-one-out
less important POD modes may exhibit broad-banded characteris- cross-validation procedure. Given strains measured at 24 sensor
tics that which may not be accurately captured by WWA; how- locations, first, one sensor (among the original 24) is selected
ever, errors introduced by ignoring these modes are small because as the target sensor; its measurement is hidden and not
only a small fraction of energy is carried by those modes. The employed in the response reconstruction. The remaining 23
effectiveness of the empirical fatigue damage estimation proce- measurements are used as inputs to the empirical method to
dure is demonstrated by applying it to the NDP riser VIV experi- reconstruct the risers response over its entire span (including
mental data sets [8]. Shortcomings of the proposed procedure are at the location of the target sensor). The reconstructed response
also discussed. at the target sensor location is then compared with the direct
measurement at the same location (which was never used for
the reconstruction). As part of the procedure, each of the avail-
able sensors is selected in turn as the target sensor in the
Model Riser and Data Sets cross-validation exercise. A parameter, termed damage ratio
The NDP model riser experiments were carried out by the Nor- (DR), defined as the ratio of the fatigue damage computed
wegian Marine Technology Research Institute (Marintek) in 2003, using the reconstructed response to the fatigue damage com-
by horizontally towing a flexible cylinder which, in combination puted using the true measurement at the same location, is then
with the actual flow conditions, can adequately describe a vertical derived as follows:
straight riser in the ocean basin test facility. The cylinder is 38 m
long with a length-to-diameter (L/D) ratio of 1400. The model ne Sbe
riser was tested for uniform as well as linearly-sheared current DR (1)
nm Sbm
profiles. Riser response was measured using 24 strain sensors (one
sensor failed for some test runs; hence, in some cases, only 23 where ne and Se represent, respectively, the number of cycles
strain sensors were available) and eight accelerometers for the and the equivalent stress range of the reconstructed stress time
cross-flow (CF) direction. Similarly, in the in-line (IL) direction, series, while nm and Sm are the corresponding quantities of the
measurements from 40 strain sensors and eight accelerometers measured stress time series at the same location. Also, b is the
were available. Key parameters of the model riser are provided in Wohler exponent of the fatigue S-N curve. In this study,
Table 1. Additional details on the NDP experiments that relate to fatigue damage is computed using the rainflow cycle-counting
the test setup and other physical properties of the model riser may algorithm and Miners rule with an assumed value of b 3.
be found in the work of Braaten and Lie [8] and Trim et al. [3]. Note that if DR < 1, the empirical reconstruction procedure
Among the six data sets available at the VIV data repository underestimates the risers response in terms of the fatigue
[9], four of them were obtained from tests on bare risers. Table 2 damage, while DR > 1 means the empirical method overesti-
summarizes current speed characteristics and root-mean-square mates it. Note too that compared with the use of, say, the riser
(RMS) values of CF displacement normalized with respect to the RMS response, the parameter, DR, is a more robust indicator
diameter, D, of the cylinder. Because of the large RMS displace- to assess the quality of the response reconstruction because it
ment values that were computed from data in these four tests, this conveys information both on the amplitude (stress range) and
response is thought to be associated with VIV and, as such, these on the frequency (number of cycles). Furthermore, DR is a
more properly discriminating indicator of damage because
errors in the stress range are magnified by a power of b. For
Table 1 Physical and hydrodynamic properties of the NDP example, if the stress range is overestimated by a factor of 2,
model riser with a S-N curve where b 3, what would result is that
DR 8, which means that the estimated fatigue damage com-
Length (m) 38 puted using the reconstructed response is eight times the true
Outer diameter (mm) 27 value.
Wall thickness (mm) 3
In the following, a sheared current data set, NDP 2350, with
Mass of riser per unit length (kg/m) 0.933
Mass of displaced water per unit length (kg/m) 0.576 strain sensor no. 4 (at a location where z/L 0.11; L is the risers
Mean effective tension (N) 40006000 length) selected as the target sensor, is employed to illustrate the
Bending stiffness (Nm2) 598.8 details of the proposed empirical method. Locations of the 23
Maximum Reynolds number 70,000 input sensors (circle symbols) and the assumed target sensor
(asterisk symbol) are indicated in Fig. 1.

031701-2 / Vol. 139, JUNE 2017 Transactions of the ASME

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or


Fig. 1 Locations of the 23 input sensors and one target sensor
(sensor no. 4) used to illustrate the empirical procedure. Note
pinned-pinned boundary conditions are in effect at x/L 5 0 and
x/L 5 1.

Results obtained from using all the four available NDP data
sets with other sensors selected as the target are also presented in Fig. 2 Normalized PSD functions of the 23 input (measured)
the Empirical Fatigue Damage Estimation Results section. strain time series

Proper Orthogonal Decomposition


Proper orthogonal decomposition (POD) is a technique widely
used in many scientific disciplines, such as oceanography, meteor- energy is concentrated around 3.5 Hz, which is the vortex shed-
ology, statistics, etc., often, with different names such as empirical ding frequency; a relatively small fraction of energy is also evi-
orthogonal function (EOF) analysis, principal component analysis dent at frequencies around 10 Hz and 16 Hz, which are associated
(PCA), and singular value decomposition (SVD). The method is with three times and five times the vortex shedding frequency,
useful for extracting energetic spatial modes or patterns of vari- respectively.
ation of any physical phenomenon that is fully described by a Following the procedure outlined above, 23 POD mode shapes,
high-dimensional spatio-temporal stochastic field (such as the /j , can be obtained; the original 23 strain time series can be
suite of riser strain time series from multiple sensors that we have decomposed into 23 uncorrelated POD scalar subprocesses, uj(t).
here). Examples of the application of POD for the analysis of riser If ranked by energy (eigenvalue), the first three POD modes are
VIV response may be found in the literature (see, for example, seen to account for 45.6%, 33.2%, and 12.2% of the total energy,
Kleiven [2], Srivilairit et al. [10], and Srivilairit and Manuel [11]). respectively, while the remaining 20 POD modes account for less
Given a suite of strain time series measured at M locations, than 10% of the total energy. In the following, the first 13 POD
Vt fv1 t; v2 t; ; vM tgT , one can establish an M  M modes, which represent 99% of the total energy, are retained for
covariance matrix, Cv, from the strain time series, V(t). By solv- response reconstruction, i.e., N 13 in Eq. (4); the remaining ten
ing an eigenvalue problem, one can diagonalize Cv so as to obtain modes (accounting for less than 1% of the energy) are assumed as
the matrix, K, as follows: noise and can be discarded. It needs to be emphasized that POD is
very efficient in extracting energetic vibration modes and the
UT Cv U K; Cv U UK (2) method effectively cleans the data; however, the mode shapes are
only defined at the locations of the input sensors. By using Eq.
(4) to reconstruct the riser response at locations where no sen-
Solution of the eigenvalue problem yields eigenvalues,
sor is available, mode shape coordinates at those locations
fk1 ; k2 ; ; kM g diagfKg (where k1 > k2 > > kM ) and asso-
need to be estimated. In a previous study by Shi et al. [7],
ciated eigenvectors, U f/1 ; /2 ; ; /M g.
third-order polynomials were used to interpolate between the
It is possible to rewrite the original M correlated time series,
discrete POD mode shape coordinates empirically available
V(t), in terms of uncorrelated scalar subprocesses,
only at sensor locations; this makes the response reconstruction
Ut fu1 t; u2 t; ; uM tgT , such that
method based on POD appear to be an ad hoc one. To address
X
M this issue, in the present study, spatially continuous mode
Vt UUt /j uj t (3) shapes are estimated by employing WWA to each of the dis-
j1 crete POD mode shapes; this improves the objectivity of the
empirical method.
where /j represents the jth POD mode shape associated with the Power spectra are estimated for the first 13 POD subpro-
jth scalar subprocess, uj(t). The energy associated with uj(t) is cesses, uj(t), j 1 to 13, and normalized to reveal the fre-
described in terms of the corresponding eigenvalue, kj. A quency content of each subprocess by setting the maximum
^
reduced-order representation of the strain time series, Vt, may PSD ordinate to unity. As illustrated in Fig. 3, the first three
be obtained by including only the first N POD modes and associ- POD subprocesses resemble narrow-banded single-peak random
ated generalized subprocesses processes, with peaks around 3.5 Hz, 3.5 Hz, and 10 Hz, respec-
tively. Given a specific number of input sensors, WWA is
X
N more efficient when applied to narrow-banded single-frequency
^
Vt /j uj t; N<M (4) processes versus broad-banded multifrequency processes; these
j1 hints at the possibility of using WWA to efficiently further
decompose POD modes and reconstruct response time series
For the data set, NDP 2350, strains were measured at 24 loca- using continuous modes aided by both POD and WWA in tan-
tions. With sensor no. 4 selected as the target, the remaining 23 dem. The decomposition and reconstruction using WWA may
strain time series are used as inputs, and thus, M 23. Normalized be expected to have high accuracy when applied to dominant
power spectral density (PSD) functions of the 23 input strain time single-frequency POD modes (recall, in our illustration, that
series are presented in Fig. 2. Instead of comparing the energy car- the first three modes account for over 90% of the total
ried by each sensor, the purpose here is to assess the frequency energy). The accuracy will be lower for the other multifre-
content of each strain time series; hence, each PSD function is quency modes; however, errors introduced by these lower
normalized so that the maximum value equals unity. Figure 2 modes to the total field reconstruction will be small because of
shows that, for all input (measured) strains, a large fraction of the relatively smaller fraction of energy associated with them.

Journal of Offshore Mechanics and Arctic Engineering JUNE 2017, Vol. 139 / 031701-3

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or


2 3 2 3 2 3
u001
1 u001
2  u001
N w1 e1 =R
6 7 6 7 6 7
6 002 7 6 7 6 7
6 u1
6 u002
2  u002
N 7
7 6 w2 7
6 7
6 e2 =R 7
6 7
A6 7; w6 7; d6 7
6 7 6 7 6 7
6 7 6 7 6 7
4 5 4 5 4 5
u00M
1 u00M
2
00M
   uN wN eM =R
(8)

where u00j i is the curvature of the nith mode shape at logger loca-
tion, zj, and ej e(zj, t) is the strain measured at location, zj. Equa-
tion (7) represents a linear system of M equations with N weights
to be estimated. At any instant of time, t, as long as N  M, the
modal weights vector, w, may be solved for in a least-squares
Fig. 3 Normalized PSD functions for the first 13 POD sense. Thus, we have
subprocesses
wt AT A1 AT dt (9)

The risers response, such as its CF bending strain, at any location,


Weighted Waveform Analysis z, can be reconstructed using Eq. (6), once the modal weights
Weighted waveform analysis (WWA) is a useful tool to repre- have been estimated.
sent the participation of different modes of vibration of a riser The WWA method, as presented earlier, appears to be fairly
toward its overall response [3,12]. In previous studies by Shi et al. simple and straightforward. However, a successful application of
[4,13], WWA was employed to decompose direct response meas- WWA has two requirements: (a) definition of the mode shapes;
urements, and then, to reconstruct the response over the entire and (b) selection of an optimal mode set. These are discussed in
riser span. In general, a risers VIV response is a multifrequency the Mode Shape Definition section.
process, which needs a sufficient number of sensors to be used
with WWA in order to represent all the important participating
modes; the accuracy of response reconstruction is low when there Mode Shape Definition. For vertical or catenary risers, the dif-
is a small number of available sensors. It is shown above that ferential equation governing free vibration may be solved using
dominant POD modes often resemble narrow-banded single- analytical methods with eigenmodes expressed in terms of Bessel-
frequency processes. Hence, in this study, instead of applying like functions (see Pesce et al. [14], and Mazzilli et al. [15]). The
WWA to the measurements directly, it is used to extract the par- selected NDP model riser has a constant tension and uniformly
ticipating vibratory POD modes first derived empirically from the distributed physical properties along its length. The nith natural
data. mode shape representing displacement in a standing wave can be
For the sake of completeness, the mathematical derivation of approximated by a sinusoidal function, i.e.
the WWA procedure is briefly presented here.  
ni pz
Assume that the riser CF displacement, x, at location, z, and ui z sin (10)
time, t, may be expressed as a weighted sum of N assumed modes. L
Thus, we have
where L is the length of the riser. Accordingly, the curvature of
X
N the nith natural mode shape can be expressed as
xz; t wi tui z (5)
i1  2  
ni p ni pz
u00i z  sin (11)
where it is assumed that by using N (not necessarily sequentially L L
consecutive) modes, one can approximately represent the riser
displacement time series at any location, z. Also, ui z represents
the nith mode shape, while wi(t) represents the time-varying modal The possible distortion to classical standing waves arising from
weight to be applied to the nith mode shape associated with index traveling waves, often observed in riser VIV, results in more com-
i in Eq. (5). plex modes than are indicated by Eq. (10). In order to capture pos-
The bending strain, e(z, t), which is the product of the riser sible traveling wave behavior, a cosine term may be introduced as
radius, R, and the local curvature, x00 , may be expressed as a complement to the sine term for each natural mode in Eq. (10)
follows: [7]. In the context of modal analysis, the sine term is the real part
of a complex mode and is referred to as the normal shape. The
X
N cosine term is the corresponding imaginary part and is termed the
ez; t Rx00 z; t Rwi tu00i z (6) Hilbert shape, because it describes the sine term out of phase by
i1 90 deg. Note that the cosine terms do not satisfy the NDP model
riser pinned-pinned boundary conditions where the transverse
where the symbol 00 denotes the second spatial derivative; so, response is zero. To correctly address this, the Hilbert shapes are
u00i z represents the curvature of the nith mode shape. adjusted spatially, using a window that decays to zero at the boun-
Given strain measurements or, equivalently, curvature measure- daries [6]. Figure 4 shows an example of normal and Hilbert shape
ments at M logger locations, zj (where j 1 to M), WWA requires components for the fifth natural mode of the NDP model riser that
solution of a system of equations in matrix form allows for possible traveling wave effects.
Generally, for a marine riser, the normal mode shapes can be
Aw d (7) estimated using the finite element method. Hilbert shapes can be
obtained by applying a Hilbert transform and a decaying spatial
where the matrix, A, comprises curvatures of the assumed mode window adjustment to the normal shapes at the extremities. With
shapes at all the logger locations and the vector, d, is formed from the complementary Hilbert shapes added to the normal shapes, the
the measured strains at all loggers. In expanded form, we have following mode shape matrix, A, results:

031701-4 / Vol. 139, JUNE 2017 Transactions of the ASME

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or


subprocess, that mode is classified as a single-frequency, double-
frequency, or triple-frequency process. The procedure for prepar-
ing the candidate mode sets is different for the three classes.
(a) If the ratio of the second spectral peak to the first peak is
less than a specified value, e.g., 0.1 (which means that the
height of the second peak is less than 10% of the first
peak), the POD subprocess is classified as a single-
frequency process. If the identified mode at the first peak is
the kth natural mode, three candidate mode sets, each con-
taining three consecutive modes, are prepared, which are
the set of the [(k  2)th, (k  1)th, kth], [(k  1)th, kth,
(k 1)th], [kth, (k 1)th, (k 2)th] modes. The reason for
considering several candidate mode sets with modes shifted
around the peak mode is to recognize that the peak mode
may not be accurately identified due to the approximate
estimation of the risers natural frequencies.
(b) If the ratio of the second spectral peak to the first peak is
greater than 0.1, while the ratio of the third to the first peak
is less than 0.1, the POD subprocess is classified as a
double-frequency process. If the modes associated with the
first and second spectral peaks are the kth and pth natural
modes, six candidate mode sets are prepared, which are the
set of the [(k  1)th, kth, (p  1)th], [(k  1)th, kth, pth],
Fig. 4 Normal and Hilbert shape components of the fifth natu- [(k  1)th, kth, (p 1)th], [kth, (k 1)th, (p  1)th], [kth,
ral mode of the NDP model riser (with pinned-pinned boundary (k 1)th, pth], [kth, (k 1)th, (p 1)th] modes. Note that
conditions) two consecutive modes are used to capture the first peak
mode because it carries a larger portion of energy compared
2 3 with the second peak mode.
u001
1 w001
1 u001
2 w001
2  u001
N w001
N (c) if the ratio of the third spectral peak to the first peak is
6 7
6 u002 w002 u002 w002  u002 w002 7 greater than 0.1, the POD subprocess is classified as a
6 1 1 2 2 N N 7
A6 7 (12) triple-frequency process; then, all the three peaks need to
6 7
4 5 be included in the WWA scheme. If the first spectral peak
u00M w00M u00M w00M  u00M w00M is identified as the kth natural mode, a mode set containing
1 1 2 2 N N
three consecutive modes centered at the peak mode, i.e.,
where u00j 00j the [(k  1)th, kth, (k 1)th] natural modes, is considered.
i and wi are curvatures of the normal mode and the Hil-
bert mode, respectively, of the nith mode shape curvature at log- This set is assumed a mode pool for the first peak; simi-
ger location, zj. larly, if the second and third peaks are identified as the pth
and rth natural modes, the associated mode pools are
[(p  1)th, pth, (p 1)th] and [(r  1)th, rth, (r 1)th],
Mode Set Selection. Another issue that needs to be addressed
respectively. The candidate mode set for this triple-
has to do with selection of the set of modes to include in the
frequency process is obtained by drawing one mode from
WWA approach. As already stated, the number of available sen-
each mode pool to represent each peak; considering all per-
sors, M, limits the number of modes, N, that may be included in
mutations, 27 candidate mode sets are obtained for any
the WWA scheme, i.e., N  M; this indicates that the availability
triple-frequency process.
of a larger number of sensors allows a greater number of modes to
be included, which will generally lead to a more accurate recon- Step 3: For each empirically derived POD mode, the mode set,
struction of the riser response. For the cross-validation exercise among all candidate mode sets, that results in the lowest recon-
using the NDP data sets, 23 input sensors are available which struction error (i.e., the smallest error between the reconstructed
allows a fairly large number of modes to be included. However, in modal coordinates and the original POD modal coordinates at all
this study, only three modes, each consisting of a normal shape input sensor locations) is selected as the optimal mode set. This
and a Hilbert shape, are used in the WWA scheme. The reasons optimal mode set is then used with the WWA scheme to decom-
for selecting only three modes, instead of a larger number, are pose and reconstruct the POD mode shape. It should be noted that
that: (a) often, the risers dominant response contribution is at the as distinct from applying WWA directly to the direct response
vortex shedding frequency as well as at three and five times that measurements, where a time-varied modal weight vector, w(t), is
frequency (as seen in Fig. 2); this requires at least three modes to obtained by least-squares fitting (see Eq. (9)) in each time-step, in
capture those frequency components and (b) the use of a smaller the present study, WWA is used to decompose and reconstruct
number of modes with the WWA scheme highlights its applica- each spatial POD mode shape, /j , which is already separated
tion to more common situations where only a small number of from the temporal subprocess, uj(t) by the POD procedure (see
sensors is available. A procedure for selecting the three modes for Eq. (4)). In other words, the matrix operations, including least-
WWA to reconstruct each POD mode is discussed next. squares fitting, are performed only once for each POD mode
Step 1: Calculate the risers natural frequencies using the ten- shape; this results in a modal weight vector, w, that is independent
sioned beam equation. For a riser with complex geometry and of time.
physical properties, the natural frequencies may be estimated Illustration: The first three dominant POD modes, which are
using the finite element method; single-frequency processes, are used as examples and presented in
Step 2: For each POD mode, a spectral analysis is performed on Fig. 5. Normalized PSDs of the scalar subprocesses associated
the associated POD subprocess. The three most dominant spectral with the POD modes are plotted at the top; these clearly show
peaks are identified and sorted based on the PSD ordinates; associ- dominant narrow-banded single spectral peaks in each case. The
ated mode numbers are identified by matching the peak frequen- POD mode shapes are plotted below, where the blue circles repre-
cies with the nearest riser natural frequencies. By comparing the sent the discrete mode shape coordinates at the sensor locations,
relative heights of the three spectral peaks for each POD mode and the red solid lines indicate the continuous mode shape

Journal of Offshore Mechanics and Arctic Engineering JUNE 2017, Vol. 139 / 031701-5

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or


Fig. 7 Strain time series at a selected sensor location (z/
L 5 0.11): direct measurements versus reconstruction
Fig. 5 Normalized PSDs and mode shapes of the first three
dominant POD modes that account for 45.6%, 33.2%, and 12.2% over its entire span, including the target location (i.e., z/L 0.11
of the total energy, respectively in this example), is estimated by combining all the important POD
modes using Eq. (4). In Fig. 7, strain time series over a 6 s time
window are plotted; results show that the reconstructed time series
functions reconstructed using WWA. The percent of energy car- (red dashed line) matches the direct measurements (blue solid
ried by each POD mode is also indicated in the figure. Figure 5 line) reasonably well. The estimated fatigue damage and the
shows that the reconstructed mode shape functions accurately true fatigue damage are computed using the reconstructed strain
describe the POD mode shape coordinates for the first three POD time series and the direct measurements, respectively, at the
modes, which together account for over 90% of the total energy. selected locations. The damage ratio, defined in Eq. (1) as the
In a similar manner, the fourth and the seventh POD modes are ratio between the estimated fatigue damage and the true fatigue
presented in Fig. 6. The fourth mode, which accounts for about damage, is 1.05, which means that the proposed empirical method
1.8% of the total energy, is a double-frequency process, while the overestimates the fatigue damage rate by 5%.
seventh mode, which accounts for about 0.9% of the total energy, Figure 8 presents RMS values of the CF strains over the entire
is a triple-frequency process. Note that the accuracy of mode span of the riser. Blue circles represent RMS values at the input
shape reconstruction for the multiple-frequency processes is not sensor locations, the red line represents corresponding values
as good as that for the single-frequency processes (seen in Fig. 5). based on the reconstruction approach, and the blue and red aster-
This is expected because it is generally more difficult to describe isks are the measured and estimated values, respectively, at the
a multifrequency process than it is to describe a single-frequency target sensor location. It is seen that the reconstruction approach
process with the same number of modes. Nevertheless, errors captures the risers global response quite accurately.
introduced in these multi-frequency modes toward the final recon-
struction are relatively small because only a small fraction of
energy is associated with these modes. Empirical Fatigue Damage Estimation Results
After a continuous mode shape function is reconstructed for In the preceding sections, a sheared current data set, NDP 2350,
each of the important POD modes using WWA, the riser response with strain sensor no. 4 (located at z/L 0.11, where L is the riser
length) selected as the target sensor, was used to demonstrate the
empirical procedures proposed for fatigue damage estimation.
Results for all four NDP riser data sets with other sensors selected

Fig. 6 Normalized PSDs and modal shapes of the fourth and


seventh POD modes that account for 1.8% and 0.9% of the total
energy, respectively Fig. 8 RMS values of CF strains over the entire riser span

031701-6 / Vol. 139, JUNE 2017 Transactions of the ASME

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or


Fig. 9 Fatigue damage ratio estimated using the proposed empirical method
(POD combined with WWA) with 23 strains as input. The range of damage ratios
estimated with each data set is indicated in the legend.

Fig. 10 Fatigue damage ratio estimated using the POD method (with interpolation
using third-order polynomials) with 23 strains as input (an earlier study by Shi
et al. [7]). The range of damage ratios estimated with each data set is indicated in
the legend.

as the target are compiled together in Fig. 9. Fatigue damage empirically derived POD mode shapes are spatially discrete func-
ratios estimated at each of the 24 sensor locations for the two uni- tions that are defined only at the sensor locations. In order to esti-
form current data sets (red color) and the two sheared current data mate the riser response as well as fatigue damage at locations
sets (blue color) are presented. The lowest and highest values of where no sensors are available, each POD mode shape is further
the fatigue damage ratio estimated for each data set are also indi- decomposed using a WWA scheme with continuous mode shape
cated in the figure legend. Considering all the four data sets, the functions. In the WWA scheme, a complementary pair of normal
fatigue damage ratio estimated using the proposed procedure with and windowed Hilbert mode shape are used together to describe
23 input sensors ranges from 0.20 to 4.75, which means that the both standing and traveling wave behavior. A procedure for
fatigue damage is, at worst, underestimated or overestimated by a selecting an optimal mode set for the WWA scheme is presented.
factor of five. The empirical procedures are demonstrated with an example study
For the sake of comparison, fatigue damage ratios estimated involving a sheared current data set, NDP 2350, with 23 input
using the original POD method, with discrete POD modes interpo- strains. Results from analysis with four available data sets show
lated using a third-order polynomial, are presented in Fig. 10 that fatigue damage computed using the empirically reconstructed
using data from a previous study by Shi et al. [7]. For the same riser response is quite accurate compared with the true value
four data sets with the same 23 input sensors, the fatigue damage based on direct measurements. The finding from this study suggest
ratio ranges from 0.17 to 12.90. The comparison suggests that the that the proposed empirical method can be used effectively in
accuracy in fatigue damage estimation is greatly improved by the reconstructing a risers response, and then, to estimate the fatigue
new empirical method outlined in the present study. damage over the entire span, using measurements at a limited
number of sensors.
Conclusions
Acknowledgment
In this study, a new procedure is developed for the purpose of
riser fatigue damage estimation, in which two modal decomposi- Financial supports provided to the first author by the National
tion techniques, POD and WWA, are sequentially applied to Key Research and Development Program Research on nondes-
response measurements. Specifically, given a series of riser tructive testing, intelligent monitoring, and experimental verifica-
response measurements, POD is first employed to decompose the tion of a multipurpose flexible pipe for ultra deepwater
original data into a set of POD modes, each expressed as a product applications (Grant No. 2016YFC0303803) and the Fundamental
of a spatial vibration pattern (mode shape) and a temporal func- Research Funds for Central University (Grant No. 16CX02024A)
tion termed a scalar subprocess. Usually a small number of POD are greatly appreciated. The authors acknowledge with gratitude
modes accounts for a large portion of the total energy, while the the permission granted by the Norwegian Deepwater Programme
remaining modes, that account for only a small fraction of the Riser and Mooring Project to use the riser high mode VIV tests.
energy, may be assumed to less important and their contribution
may then be ignored. The purpose for using POD is, as such, to References
clean the data and, more importantly, to extract energetic mode [1] Tognarelli, M., Taggart, S., and Campbell, M., 2008, Actual VIV Fatigue
contributions directly from the data. A problem that arises is that Response of Full Scale Drilling Risers: With and Without Suppression

Journal of Offshore Mechanics and Arctic Engineering JUNE 2017, Vol. 139 / 031701-7

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or


Devices, 27th International Conference on Offshore Mechanics and Arctic [9] MIT, 2007, Vortex Induced Vibration Data Repository, Center for Ocean
Engineering (OMAE2008), Estoril, Portugal, June 1520. Engineering, Massachusetts Institute of Technology, Cambridge, MA, http://
[2] Kleiven, G., 2002, Identifying VIV Vibration Modes by Use of the web.mit.edu/towtank/www/vivdr/datasets.html
Empirical Orthogonal Functions Technique, ASME Paper No. OMAE2002- [10] Srivilairit, T., Manuel, L., and Saranyasoontorn, K., 2006, On the Use of
28425. Empirical Orthogonal Decomposition of Field Data to Study Deepwater Current
[3] Trim, A., Braaten, H., Lie, H., and Tognarelli, M., 2005, Experimental Investi- Profiles for Risers, 16th International Offshore and Polar Engineering Confer-
gation of Vortex-Induced Vibration of Long Marine Risers, J. Fluids Struct., ence (ISOPE), San Francisco, CA, May 28June 2, pp. 136140.
21(3), pp. 335361. [11] Srivilairit, T., and Manuel, L., 2009, Vortex-Induced Vibration and Coincident
[4] Shi, C., Manuel, L., Tognarelli, M., and Botros, T., 2012, On the Vortex- Current Velocity Profiles for a Deepwater Drilling Riser, ASME J. Offshore
Induced Vibration Response of a Model Riser and Location of Sensors for Mech. Arct. Eng., 131(2), p. 021101.
Fatigue Damage Prediction, ASME J. Offshore Mech. Arct. Eng., 134(3), [12] Lie, H., and Kaasen, K., 2006, Modal Analysis of Measurements From a
p. 031802. Large-Scale VIV Model Test of a Riser in Linearly Sheared Flow, J. Fluids
[5] Lucor, D., and Triantafyllou, M., 2008, Riser Response Analysis by Modal Struct., 22(4), pp. 557575.
Phase Reconstruction, ASME J. Offshore Mech. Arct. Eng., 130(1), p. 011008. [13] Shi, C., Park, J., Manuel, L., and Tognarelli, M., 2014, A Data-Driven Mode
[6] McNeill, S., and Agarwal, P., 2011, Efficient Modal Decomposition and Identification Algorithm for Riser Fatigue Damage Assessment, ASME J. Off-
Reconstruction of Riser Response Due to VIV, ASME Paper No. OMAE2011- shore Mech. Arct. Eng., 136(3), p. 031702.
49469. [14] Pesce, C., Fujarra, A., Simos, A., and Tannuri, E., 1999, Analytical and Closed
[7] Shi, C., Manuel, L., and Tognarelli, M., 2010, Alternative Empirical Proce- Form Solutions for Deep Water Riser-Like Eigenvalue Problem, 9th Interna-
dures for Fatigue Damage Rate Estimation of Instrumented Risers Undergoing tional Offshore and Polar Engineering Conference (ISOPE-1999), Brest,
Vortex-Induced Vibration, ASME Paper No. OMAE2010-20992. France, May 30June 4, pp. 255264.
[8] Braaten, H., and Lie, H., 2004, NDP Riser High Mode VIV Tests, Technical [15] Mazzilli, C. E., Rizza, F., and Dias, T., 2016, Heave-Imposed Motion in Verti-
Report, Norwegian Marine Technology Research Institute, Trondheim, Nor- cal Risers: A Reduced-Order Modelling Based on Bessel-Like Modes, Proce-
way, Main Report No. 512394.00.01. dia IUTAM, 19, pp. 136143.

031701-8 / Vol. 139, JUNE 2017 Transactions of the ASME

Downloaded From: http://offshoremechanics.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmoeex/936068/ on 02/17/2017 Terms of Use: http://www.asme.or

You might also like